{"id":"76af6fc7-abc0-4c56-8302-21ae9010487b","arxiv_id":"2507.06247","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A 350GB synthetic dataset pairing frame and event recordings of particles with ground-truth flow fields, plus a reusable generator, is introduced.","lead":"FED-PV is a new synthetic dataset that pairs ordinary particle frames with event-camera signals for nine flow types, complete with velocity ground truth. It is meant to give particle-velocimetry researchers a shared resource for building and benchmarking algorithms that fuse both camera modalities.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 2 in Sec. 2.3 does not specify the reference time; if implemented as per-frame differencing, sub-threshold log-brightness changes never accumulate, making slow-flow event streams unrepresentative of real event cameras.","rationale":"The dataset's value is as a surrogate for real frame/event recordings; the event stream is half of the dual-modal promise. A per-frame threshold implementation would make event data for low-velocity flows qualitatively different from any event camera, so algorithms developed on FED-PV (and the Table 3 benchmarks) would not transfer. The reader's weakest assumption identified event-camera realism broadly; my concern is a sharper, checkable sub-case concerning the thresholding reference time. I do not think the paper is dishonest; the ambiguity may simply be under-specified text. Because the paper references ESIM and uses 100-µs interpolation, the implementation could well be standard, so this is a conditional concern rather than a demonstrated error. The existing CONDITIONAL verdict is appropriate: the release of the generator and a clear statement of the event-generation logic would resolve it. Thus I recommend no change to the reader's verdict.","tokens_in":7143,"tokens_out":15037,"duration_ms":170077,"concrete_test":"Obtain the event-generation routine from the released repository; if the repository is unavailable, request the code from the authors. Run it on a single Gaussian particle spot (dp ~3 px) moving at 0.1 px/ms for 15 ms with threshold C=0.25 and 100-µs rendering. Count events and compare with ESIM run on the same input. If FED-PV yields zero events while ESIM produces cumulative-threshold events, per-frame differencing is confirmed and the slow-velocity subsets are unrealistic; if counts and timestamps match ESIM, the event model is standard and the concern is dismissed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. 2 in Sec. 2.3 defines an event by |log I(x,y,t1) - log I(x,y,t2)| >= C, but the reference time t2 is not defined as the time of the last event at that pixel. If the generator checks adjacent 100-µs frames only, then a log-intensity increment that is below C per frame never accumulates, so slowly advected particles (e.g., 0.1-0.2 px/ms in the DNS-turbulence and uniform subsets, Table 1) will trigger essentially no events, even though a real DVS would fire once the accumulated log-contrast since the last event reaches C. The manuscript does not state that the code uses the standard ESIM threshold-crossing method, nor does it validate event counts or timestamps against ESIM or a real event camera. Sec. 3.2 therefore benchmarks event streams whose fidelity to real event data is unestablished. If the implementation is actually ESIM-style, the concern is resolved by documentation and a reproducibility check; as written, the central claim that FED-PV provides usable cross-modal data for PV rests on an unspecified event-model detail.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces FED-PV, a dual-modal synthetic dataset generator and an associated 350 GB dataset for particle-based velocimetry (PV). The generator advects Gaussian-modeled particles through nine flow-field types taken from Cai et al. (2019), extracts event streams by thresholding log-brightness changes, retains four PIV frames per scenario, and provides ground-truth velocity fields. The paper benchmarks three frame-based PIV methods (UnLiteFlowNet-PIV, Cross-correlation-PIV, RAFT-PIV) and three event-based methods (EBIV, E-RAFT, Contrast-Maximization) to demonstrate usability. The central claim is that FED-PV fills the missing cross-modal frame/event data gap for developing fusion-based PV algorithms.","tokens_in":7412,"tokens_out":4303,"duration_ms":48469,"significance":"If the event-stream generation is faithful to real event-camera behavior, FED-PV is a potentially valuable community resource: it is large-scale, includes ground truth, covers diverse flow types, and ships an open-source generator. The benchmark across six algorithms provides a useful point of reference. The main strengths are the scale of the released dataset and the explicit synthetic-ground-truth design. However, the paper's core value depends on the fidelity of the simulated event data, and that fidelity is currently not established because the event-generation mechanism is under-specified; the benchmark metrics are also undefined. These issues are fixable in revision and do not undermine the overall utility of the dataset if addressed.","major_comments":[{"comment":"The reference time t2 in Eq. (2) is not defined. If events are generated by comparing only consecutive 100 µs frames, sub-threshold log-intensity increments never accumulate, so slow flows (e.g., the DNS-turbulence and uniform subsets with average velocities as low as 0.1-0.2 px/ms in Table 1) would trigger almost no events even though a real event camera would fire once the accumulated contrast since the last event reaches the threshold. Please specify that the reference is the last-event time at each pixel, and state whether the implementation follows ESIM's threshold-crossing scheme; a comparison of event counts or timestamps against ESIM or a real event recording would substantiate the fidelity claim.","section":"Sec. 2.3, Eq. (2)"},{"comment":"The temporal model is internally inconsistent. Section 2.2 states that PIV image sequences are generated at fixed 1 ms intervals and that the last four frames are retained, while Section 2.3 states that the event inter-frame interval is 100 µs and Fig. 3 labels a frame interval as 100 µs. Please clarify whether the event stream is synthesized from a separate high-frame-rate image stream that is not identical to the retained PIV frames, and how the 15 ms event window aligns with the timestamps of the four retained images shown in Fig. 1.","section":"Secs. 2.2, 2.3 and Figs. 1, 3"},{"comment":"The evaluation metrics RMSE, AEE, and AAE are never defined. Without formulas and units (particularly for the angular error AAE), the benchmark numbers cannot be reproduced or compared. Please also state over how many scenarios each table entry is averaged and report standard deviations or percentiles, since some rows, such as the uniform-flow row in Table 2, show very large differences among methods that would be informative only with aggregation details.","section":"Secs. 3.1-3.2, Tables 2-3"},{"comment":"The event-based benchmark is not reproducible as described. The paper does not state whether E-RAFT was used with pretrained weights or fine-tuned on the synthetic data, which event representation was fed to the network, or how the grid for Contrast-Maximization was configured. The statement that E-RAFT performs suboptimally 'due to the lack of algorithmic tuning' makes the comparison preliminary. Please report the exact configurations and clarify whether all baseline methods are used strictly off-the-shelf.","section":"Sec. 3.2"}],"minor_comments":[{"comment":"The log arguments in Eq. (2) are missing the intensity function I; the expression should read |log I(x,y,t1) - log I(x,y,t2)| >= C.","section":"Sec. 2.3, Eq. (2)"},{"comment":"The sentence explaining positive and negative events is imprecise: a negative event should be generated when the log-brightness decreases by at least C, not when the brightness change is 'below the threshold'.","section":"Sec. 2.3"},{"comment":"The axes in Fig. 3 are not labeled; the horizontal axis should indicate time with units (ms or µs) and the vertical axis should indicate log-intensity.","section":"Fig. 3"},{"comment":"The contribution bullet 'making it one dual-modal dataset available' reads awkwardly; it should say 'making it the only dual-modal dataset' or 'one of the first dual-modal datasets' as intended.","section":"Sec. 1"},{"comment":"The paper provides only a GitHub repository link; please include a persistent dataset DOI or archive link and a clear license statement so that the 350 GB release is reproducible and citable.","section":"Sec. 2.4 / Data availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is primarily a dataset description. The event-model ambiguity and the undefined benchmark metrics are the main obstacles to accepting the paper in its current form. If the authors can clarify the event-generation reference time, resolve the 1 ms / 100 µs timing inconsistency, and define the metrics, the contribution would be suitable for publication as a dataset resource."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take: FED-PV is a genuinely new resource. No one else has released synchronized frame and event particle data with ground truth for velocimetry, and the generator is customizable to flow type, particle density, and event threshold. That's a real gap, and the paper fills it. The benchmark effort with three PIV and three event-based methods is a reasonable first sanity check; the numbers in Tables 2 and 3 are plausible, and the qualitative flow reconstructions look sensible.\n\nThe main soft spot is the event model. Equation (2) says an event fires when |log I(t1) - log I(t2)| >= C, but t2 is never defined. If it's the previous 100-µs frame, then sub-threshold brightness changes never accumulate. For the slower subsets (DNS turbulence, JHTDB channel hd, uniform at the low end), particle displacements can be well under 0.1 px per 100 µs, so you'd get almost no events. That would make the event benchmarks meaningless for exactly the flows where you need them. If, instead, t2 is the time of the last event at that pixel—the standard ESIM approach—then the model is fine, but the paper has to say that and, ideally, verify the implementation against ESIM or a real event camera. The current text is ambiguous, and this is a load-bearing detail.\n\nThe other issues are smaller. The frame-timing description is easy to misread: Sec 2.2 says PIV images at 1 ms intervals, Sec 2.3 says an inter-frame interval of 100 µs. I think the 100 µs is the event-generation stream, but the paper should state that explicitly. The metrics (RMSE, AEE, AAE) are never defined. The conclusion claims 'high-fidelity' modeling, but there is no validation against physical event data, so scale that claim back to what was actually done. The GitHub link is given but there is no direct dataset link or access statement.\n\nNone of these are fatal. The event-model ambiguity is the one that matters, and it is fixable with a short clarification and a reproducibility check. The paper is worth refereeing, and once the event model is documented and verified, I'd expect this to become a standard reference for anyone working on event-based or fusion-based particle velocimetry.","headline":"Useful new multimodal dataset for particle velocimetry, but the event generation needs to specify its threshold-crossing model before I'd trust the slow-flow subsets.","tokens_in":7941,"tokens_out":4050,"would_cite":true,"duration_ms":46814,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that the lack of cross-modal datasets combining frame-based particle images and event-camera streams is the main bottleneck for next-generation particle-based velocimetry, and introduces FED-PV, a generator and 350 GB…","keywords":["particle image velocimetry","event camera","synthetic dataset","multimodal fusion","flow field measurement","event-based imaging velocimetry","deep learning"],"falsifier":"Take a real event camera and record a printed particle field moving at a known constant velocity on a translation stage, then compare the event rate, spatial distribution, and direction of events with FED-PV's generated events for the same nominal flow; if the real event rate versus contrast relationship or the distribution of inter-event times differs substantially from the 25%-threshold model, the synthetic events would not transfer to physical experiments.","tokens_in":6958,"feed_emoji":"🎥","tokens_out":5679,"duration_ms":55677,"temperature":0.7,"pith_summary":"This paper argues that the lack of cross-modal datasets combining frame-based particle images and event-camera recordings is the main bottleneck for next-generation particle-based velocimetry (PV). To fill that gap, it introduces FED-PV, a generator that synthesizes synchronized particle image sequences, event streams, and ground-truth velocity fields from nine types of flow, released as a 350 GB dataset. The paper reports benchmark results with three PIV and three event-based algorithms, showing the dataset can support both conventional and event-driven velocity estimation. If the dataset's synthetic events faithfully represent real event-camera behavior, it would give the community a reusable resource for developing and testing fusion-based PV algorithms.","feed_headline":"350GB synthetic dataset pairs particle images with event streams","feed_subtitle":"Nine flow types with ground truth give fusion-based velocimetry a benchmark to train and compare algorithms.","key_machinery":"The core machinery is the FED-PV generation pipeline, which couples a particle image generator with an event simulator. Particles are modeled as two-dimensional Gaussian intensity profiles and advected through velocity fields using uniform linear motion over $0.02\\,\\mathrm{ms}$ micro-steps with bilinear interpolation of local velocity. The computational domain is extended from $256\\times256$ to $288\\times288$ with edge sub-region repopulation to keep particle density stable, and events are produced by comparing log-brightness changes between $100\\,\\mu\\mathrm{s}$ frames against a fixed 25% threshold.","core_discovery":"The central claim of the paper is that FED-PV provides the first large-scale dual-modal dataset for particle-based velocimetry, with synchronized frame and event recordings plus exact ground truth. The dataset is generated by dispersing synthetic Gaussian particles into known CFD flow fields, advecting them with small time steps, and producing events by thresholding log-brightness changes at a 25% contrast threshold. The authors demonstrate the dataset's utility by benchmarking three frame-based PIV methods and three event-based methods, reporting RMSE, AEE, and AAE statistics that show existing algorithms can recover flow structure but still leave room for improvement.","pith_inferences":["The fixed 25% contrast threshold and noise-free event model may make FED-PV events cleaner than real sensor output; algorithms tuned on this dataset could overfit to noiseless events, so adding realistic sensor noise and threshold variability would be a natural extension.","Because the event stream is derived from the same simulation as the frames, the two modalities are perfectly registered; this makes FED-PV a strong testbed for estimating the value of fusion, but real experiments with imperfect registration may show less benefit.","A direct validation of the event model against a physical event camera on a translating particle pattern would strengthen the claim that the dataset transfers to real experiments.","The dataset's structure could be reused to generate synthetic data for other flow diagnostics, such as predicting pressure fields or doing super-resolution, by adding appropriate ground truth."],"forward_implications":["Researchers can train and evaluate multimodal fusion networks for velocimetry on a common benchmark with known ground truth.","Event-based velocimetry methods can be compared against frame-based PIV baselines under identical flow conditions, enabling head-to-head assessment.","The open-source generator allows custom datasets with different particle densities, thresholds, and velocity fields, supporting domain-specific studies.","The benchmark numbers (RMSE, AEE, AAE) provide reference points that future PV algorithms can be tested against.","The dataset's event streams enable studies of event-camera temporal resolution for flows with velocities from 0.1 to 8.3 pixels/ms."],"supporting_citations":[{"why":"Supplies the nine flow-field types and ground-truth velocity fields that the dataset is built on.","marker":"(Cai et al., 2019)"},{"why":"Defines the particle image generation standard used to synthesize grayscale PIV images.","marker":"(Raffel et al., 2018)"},{"why":"Provides the event-camera simulation principle of deriving events from image brightness changes.","marker":"(Mueggler et al., 2017)"},{"why":"Informs the event generation approach with the open ESIM event simulator.","marker":"(Rebecq et al., 2018)"},{"why":"Supplies the event-based imaging velocimetry method used to benchmark the event data.","marker":"(Willert and Klinner, 2022)"},{"why":"Provides the RAFT architecture used as the RAFT-PIV baseline in the frame-based benchmark.","marker":"(Teed and Deng, 2020)"}],"fun_headline_variants":["First dual-modal PV dataset: 350GB of frame and event pairs","Event streams meet particle images in 350GB synthetic dataset","350GB synthetic benchmark for fusion-based velocimetry algorithms","Dual-modal velocimetry dataset: 350GB of particle frames and events"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that thresholded log-brightness differences computed from synthetic Gaussian particle images with a fixed 25% contrast threshold and no sensor noise faithfully reproduce what a real event camera would record from a particle-laden flow.","fun_headline_variants_meta":{"raw":{"variants":["First dual-modal PV dataset: 350GB of frame and event pairs","Event streams meet particle images in 350GB synthetic dataset","350GB synthetic benchmark for fusion-based velocimetry algorithms","Dual-modal velocimetry dataset: 350GB of particle frames and events"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000675,"raw_usage":{"total_tokens":2995,"prompt_tokens":790,"completion_tokens":2205,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":406,"completion_tokens_details":{"reasoning_tokens":2132}},"tokens_in":406,"tokens_out":2205,"duration_ms":20877,"temperature":1.0,"reasoning_tokens":2132,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:16:56.556695+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real event camera and record a printed particle field moving at a known constant velocity on a translation stage, then compare the event rate, spatial distribution, and direction of events with FED-PV's generated events for the same nominal flow; if the real event rate versus contrast relationship or the distribution of inter-event times differs substantially from the 25%-threshold model, the synthetic events would not transfer to physical experiments.","supporting_citations":[{"cited_title":"Experiments in Fluids 60:1--16","cited_arxiv_id":null,"evidence_quote":"Supplies the nine flow-field types and ground-truth velocity fields that the dataset is built on."},{"cited_title":"springer","cited_arxiv_id":null,"evidence_quote":"Defines the particle image generation standard used to synthesize grayscale PIV images."},{"cited_title":"The International Journal of Robotics Research 36:142--149","cited_arxiv_id":null,"evidence_quote":"Provides the event-camera simulation principle of deriving events from image brightness changes."},{"cited_title":"Experiments in Fluids 63:101","cited_arxiv_id":null,"evidence_quote":"Supplies the event-based imaging velocimetry method used to benchmark the event data."},{"cited_title":"in Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part II 16","cited_arxiv_id":null,"evidence_quote":"Provides the RAFT architecture used as the RAFT-PIV baseline in the frame-based benchmark."}],"review_version":1}